/enri-thumbnails/careeropportunities1f0caf1c-a12d-479c-be7c-3c04e085c617.tmb-mega-menu.jpg?Culture=en&sfvrsn=d7261e3b_1)
/cradle-thumbnails/research-capabilities1516d0ba63aa44f0b4ee77a8c05263b2.tmb-mega-menu.jpg?Culture=en&sfvrsn=1bc94f8_1)
Visual Object Search
Visual Object Search in large-scale image and video datasets is gaining increasing traction from industry and academia in recent years. The research team led by Prof. Yuan Junsong has built an advanced visual object search system on a one million image dataset. Given a query object such as a logo, the search system is designed to efficiently retrieve all images containing the same logo from millions of images, as well as to accurately locate the object in these retrieved images (even when the logo is partially obscured or distorted). Such a search system is of great interest to many applications, including product search and recommendation, context-aware advertisement, etc. For example, the system can help the user to find a person’s favorite product in the online shopping sites by simply snapping a picture of the product through mobile phone. Moreover, the system can also enable intelligent video surveillance applications such as quickly finding a suspicious person or vehicle in surveillance videos.
Compared with the state-of-the-art methods, the developed system achieves better performance in both accuracy and efficiency, since it can robustly and efficiently identify the query object in images/videos with cluttered backgrounds, and is easy to be parallelized. A visual object search system, including a server-based search engine and an Android application, has been developed for mobile applications and demonstrated at ACM Conf. on Multimedia 2012. The team has also filed a provisional U.S. patent on this technology.
Visual Object Search targets on the search for planar or rigid objects such as the logos or cars. The researchers have developed a technique which takes advantage of the spatial context instead of matching individual features separately. The idea is to bundle a group of co-located features into a visual phrase based on a random partitioned window. Robust local matching can then be achieved by aggregating the matching scores over a collection of visual phrases of various sizes and shape, which is the advantage of this technique.